Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Visually Dehallucinative Instruction Generation

Sungguk Cha, Jusung Lee, Younghyun Lee, Cheoljong Yang

TL;DR

CAP2QA addresses the problem of visual hallucination in synthetic visual instructions by generating image-aligned QA data constrained to image content. The approach uses GPT-3.5/4 with image-aligned captions and a prompting scheme to ensure alignment and filter artifacts, producing CAP2QA-COCO as a large-scale dataset. Empirical results show substantial reductions in hallucination and improved zero-shot visual recognition and expressiveness compared with LLaVA baselines, across VQA and caption tasks. The work demonstrates the value of image-aligned instructed QA for safer, more reliable vision-language models and suggests avenues for scaling with web-scale caption data.

Abstract

In recent years, synthetic visual instructions by generative language model have demonstrated plausible text generation performance on the visual question-answering tasks. However, challenges persist in the hallucination of generative language models, i.e., the generated image-text data contains unintended contents. This paper presents a novel and scalable method for generating visually dehallucinative instructions, dubbed CAP2QA, that constrains the scope to only image contents. Our key contributions lie in introducing image-aligned instructive QA dataset CAP2QA-COCO and its scalable recipe. In our experiments, we compare synthetic visual instruction datasets that share the same source data by visual instruction tuning and conduct general visual recognition tasks. It shows that our proposed method significantly reduces visual hallucination while consistently improving visual recognition ability and expressiveness.

Visually Dehallucinative Instruction Generation

TL;DR

CAP2QA addresses the problem of visual hallucination in synthetic visual instructions by generating image-aligned QA data constrained to image content. The approach uses GPT-3.5/4 with image-aligned captions and a prompting scheme to ensure alignment and filter artifacts, producing CAP2QA-COCO as a large-scale dataset. Empirical results show substantial reductions in hallucination and improved zero-shot visual recognition and expressiveness compared with LLaVA baselines, across VQA and caption tasks. The work demonstrates the value of image-aligned instructed QA for safer, more reliable vision-language models and suggests avenues for scaling with web-scale caption data.

Abstract

In recent years, synthetic visual instructions by generative language model have demonstrated plausible text generation performance on the visual question-answering tasks. However, challenges persist in the hallucination of generative language models, i.e., the generated image-text data contains unintended contents. This paper presents a novel and scalable method for generating visually dehallucinative instructions, dubbed CAP2QA, that constrains the scope to only image contents. Our key contributions lie in introducing image-aligned instructive QA dataset CAP2QA-COCO and its scalable recipe. In our experiments, we compare synthetic visual instruction datasets that share the same source data by visual instruction tuning and conduct general visual recognition tasks. It shows that our proposed method significantly reduces visual hallucination while consistently improving visual recognition ability and expressiveness.
Paper Structure (8 sections, 2 figures, 3 tables, 2 algorithms)

This paper contains 8 sections, 2 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. CAP2QA
  4. Experimental Results
  5. Implementation Detail
  6. Evaluation
  7. Results
  8. Conclusion

Figures (2)

  • Figure 1: Types and samples of image-text data. Image-alignness indicates if the context of QA is relevant to the image content. Instructive QA refers to sentence-level-answer data. Difficulty level regards visual question answering.
  • Figure 2: Qualitative results in visual understanding.